AI is delivering on its promise, and that's the problem.
If one unit of development now gets done five times faster, how do we deal with the five units of alignment and management we have to do?
The Harness
My calendar used to have gaps. Not a lot (I was still busy) but there were these pockets where I'd put on headphones, open VS Code, and disappear into a problem for a few hours. That was the good stuff. That state of flow where you're holding a whole system in your head and reshaping it. I looked forward to those gaps.
My calendar doesn't have gaps anymore. Not because I'm working more hours. But something shifted in what fills those hours, and it happened so gradually I didn't notice until I was already in it.
5x faster = 1.5x faster
AI got good at the things I used to spend my day doing. Using Claude, Cursor, Copilot (pick your favourite) I can write code faster than ever. I can analyse data in a fraction of the time. I can produce more, ship more, move through technical work at a pace that would've been nuts two years ago. If you're a dev, you've 100% felt this first hand.
On pure coding tasks, our team measured something like a 5x speed improvement. But total project time? That only improved 1.5x. That made me think. Five times faster at the work. One and a half times faster at the project. Where did the other 3.5x go?
Here's my theory: it went into our calendars.
The harness
Every piece of code that gets written creates a surrounding body of work that has nothing to do with writing code. It needs to be reviewed. It needs to be communicated. Does it align with the roadmap? Does it do what the customer actually needs? Do we tell other teams about this change? Is someone else working on something that now conflicts? What happens at 2am when it breaks and nobody else on the team understands how it works?
The harness is everything that turns code into a shipped, understood, trusted, maintainable product. Code review, planning, alignment, context-sharing, decision-making, stakeholder communication, and the slow process of making sure that ten people actually mean the same thing when they say "ready."
Here's the important part: harness work doesn't scale with the size of a change. It scales with the number of changes. A one-line config tweak and a new authentication system generate surprisingly similar amounts of coordination. Both need review, communication, sign-off, documentation. The harness is per-change, not per-line.
This is exactly why AI speed-ups inflate the harness so aggressively. AI makes each change faster, but it also makes you ship more changes per week. A feature that took a week now takes a day, so five features land where one used to. But each one still generates its own full unit of harness work. Five features, five rounds of review, five alignment conversations, five "just making sure we're on the same page" syncs.
The development sped up. The harness didn't.
AI is fast. The space between us isn't.
AI made each of us faster. But it didn't make the space between us faster.
The meetings multiplied, but not the good kind, not two people half-confused and building an idea together in real time. That kind of collaboration actually got squeezed out. What multiplied is the logistics. The distributing of information. The alignment, the re-alignment after the alignment didn't stick, the re-re-alignment after someone was on holiday when the second alignment happened. If you've worked in (or with) eterprises, you'll know what I'm talking about.
When people measure AI productivity, they measure the task: "this took four hours, now it takes 45 minutes." That's real. But nobody works in tasks. People work in weeks. You have a 40-hour work week, and coding is only a part of it. Speed up that fraction by 5x and the total improvement is bounded by the harness. That's where the 5x collapses into 1.5x, and I don't think anyone is really looking at that gap.
The job is changing underneath us
Developers are quietly, gradually, irreversibly having their jobs changed. The thing we spend most of our time doing is shifting.
Think about the arc. "Coder" was once a job title. You wrote algorithms, made things faster, more elegant, more efficient. Then came "software engineer." Now your job includes understanding patterns, owning larger systems, carrying some business context. Each step moved you a little further from pure code and a little closer to coordination. A layer of abstraction. Most coding problems became marked as 'solved', so we had to start tackling 'engineering' problems. Now, I think we're at a point where AI solves most 'engineering' problems.
AI is accelerating that shift. If the code takes 45 minutes instead of four hours, and the harness still takes the same amount of time, then the harness becomes the majority of your job. Not gradually over a decade. But now. Like, this year.
And the cruel part is that AI is doing this by making us better at the thing we actually enjoy. If you've ever seen the Neverending Story, there's a character - Bastian. And Bastian can make infinite wishes, but he loses a memory for every one he makes. We're gaining incredible productive power, and paying for it with the part of the work that made us become developers in the first place. Every prompt we bang into Claude code removes our ability to think critically about what's being implemented.
Right now this feels exciting. The novelty of AI-assisted coding is genuinely amazing. But once that wears off, what's left? A job that is mostly orchestration, communication, and alignment, done by people who chose this career because they liked building things.
Can we speed up the harness?
I want to say yes. I'm not sure I believe it.
The problem isn't tooling. The problem - in my eyes - biological. There's a hard limit on how much information a single human can interpret and feed back on. If your team used to ship one feature a week and now ships five, the code can move that fast. Your product manager cannot. Nobody can meaningfully review, prioritise, and provide feedback on five times more output without something giving. Either the quality of their judgment or their sanity, or the work itself.
And then there's context switching. When developers were slower, they worked on fewer things at once. Now we're touching more projects, more features, more systems in the same week. The penalty for switching between them is massive and well-documented. It takes time to reload the mental model of a system you haven't looked at since Tuesday. AI doesn't help with this. If anything, it makes us more detached from the systems themself.
There's also a ceiling that nobody talks about. If the person who signs off on a feature is on parental leave, or in hospital, or simply asleep in another timezone, AI can't speed up that feedback loop. The harness runs on human time, and human time doesn't compress. There will always be people in the loop who don't have the full context, and there's no tool in the world right now that can replace the slow, messy process of actually getting them up to speed.
No neat conclusion
I don't have a product to pitch or a framework to offer.
I think we're making the code faster, and slowly begging to discover that the code was never the bottleneck. The bottleneck is the harness. The human, messy, unglamorous work of making sure a team of people actually builds the right thing together. And that bottleneck is growing, quietly eating our calendars, one "quick sync" at a time.